Vision-based Detection And Recognition Of Bus Passenger

Pedestrians detection and re-identification are important parts of computer vision.They mainly distinguish and locate ‘interested’ regions from video or image sequences,and then recognize and judge the target of ‘interested’ regions again.From the perspective of non-overlapping cameras,pedestrians detection and re-identification based on computer vision and pattern recognition have attracted more and more attention.Mainly because they have a wide range of applications,especially in intelligent video surveillance,transportation,criminal investigation and other important areas.In recent years,although great progress has been made in pedestrians detection and re-identification,there are still some problems in practical aplication.For example,bus,subway and the shopping mall,due to the influence of illumination,occlusion,perspective difference,background confusion and clothing color approximation,the pedestrians detection and recognition are still challenging topics.Based on these,how to efficiently and correctly detect the bus passengers,under the influence of external factors such as illumination and occlusion,they are the main content of this paper.The specific work are as follows:First for the bus passenger detection,based on the human head and shoulder,an enhanced aggregate filtering channel feature passenger detection algorithm is proposed.Due to the large volume of bus passenger traffic,and it is vulnerable to external factors such as lightness change,occlusion and so on.In this paper,we use the multi-scale Retinex pre-treatment to eliminate the influence of illumination and noise on image processing.In addition,train and learn a model of the passengers’ head and shoulder to avoid occlusion and false detection.Using the aggregated filter channel features detection algorithm to extract different and multiple scales of images channel features.The filter banks are used to filter the bottom features,last boosted decision tree is used to train and classification.Experiments on public datasets and bus passenger datasets demonstrate the effectiveness of the proposed algorithm.Next is the re-identification of the bus passengers.Based on the improved Fisher linear vector coding descriptors and the cross-viewing perspective quadratic discriminant analysis metric learning combination,a new pedestrians re-identification algorithm is proposed for our bus passengers.For the detected bus passengers,based on the simple 7-dimensional features of image pixels,described the local features.Using the Gaussian mixure models to model our feature datas.Fisher linear vector aggregate to form the channel feature descriptors.The final feature representation descriptors are obtained on the HSV color space.Then the high dimensional characteristics of the samples are studied by low-dimensional feature subspaces and metric matrices.Therefore it makes the characteristic distance of the intra-class sample smaller than the characteristic distance of the inter-class sample.For the bus passenger head-shoulder database,based on the improved Fisher linear vector encoded descriptor andcross-view quadratic discriminant analysis measurement algorithm,the recognition rate reaches 87.33%.